Validating the Surrogate Models

Fast surrogate models stand in for expensive simulations during design exploration; the record validates each surrogate against the full model it replaces and states its error.

Fast stand-ins, honestly bounded

Exploring a design space requires thousands of evaluations, which the full physics codes cannot supply quickly. Surrogate models — fast approximations trained on full-model runs — fill that gap for the breeder (Hyperion). A surrogate is only trustworthy inside the region it was trained on and to the accuracy it was validated at, and the record states both.

# Surrogate validation (schematic)
# 1. Train surrogate s(x) on full-model samples {x_i, f(x_i)}
# 2. Hold out test points not used in training
# 3. Report error: max & RMS |s(x)-f(x)| on held-out set
# 4. Record the valid input domain; refuse extrapolation beyond it

Rules the record enforces

• Held-out test error, not training error, defines a surrogate's accuracy.
• The valid input domain is stated; extrapolation is flagged, not silent.
• Headline results are confirmed on the full model, not left to the surrogate.
• Surrogate uncertainty is included in the reported intervals.

Where surrogates stop

Surrogates accelerate exploration; they do not certify results. Every quantity that appears in the record's headline claims is checked against the full model it approximates. A surrogate that disagrees with its parent model outside tolerance is a bug to fix, and its domain of validity is part of the uncertainty budget.

This page describes a design-and-simulation study, not a built machine. Construction of the breeder (Hyperion) begins Q2 2027; first-of-a-kind first tritium is targeted near 2030. No hardware net-gain claim is made before FOAK.